[CIR][ABI][AArch64][Lowering] Support structures with padding (llvm#1118)

The title describes the purpose of the PR. It adds initial support for
structures with padding to the call convention lowering for AArch64.

I have also _initial support_ for the missing feature
[FinishLayout](https://github.com/llvm/clangir/blob/5c5d58402bebdb1e851fb055f746662d4e7eb586/clang/lib/AST/RecordLayoutBuilder.cpp#L786)
for records, and the logic is gotten from the original codegen.

Finally, I added a test for verification.

Author: bruteforceboy

clang/include/clang/CIR/MissingFeatures.h

@@ -404,9 +404,6 @@ struct MissingFeatures {
   // specs. We should make it always present.
   static bool makeTripleAlwaysPresent() { return false; }
 
-  // This Itanium bit is currently being skipped in cir.
-  static bool itaniumRecordLayoutBuilderFinishLayout() { return false; }
-
   static bool mustProgress() { return false; }
 
   static bool skipTempCopy() { return false; }

clang/lib/CIR/Dialect/Transforms/TargetLowering/LowerFunction.cpp

@@ -347,6 +347,31 @@ mlir::Value emitAddressAtOffset(LowerFunction &LF, mlir::Value addr,
   return addr;
 }
 
+/// Creates a coerced value from \param src having a type of \param ty which is
+/// a non primitive type
+mlir::Value createCoercedNonPrimitive(mlir::Value src, mlir::Type ty,
+                                      LowerFunction &LF) {
+  if (auto load = mlir::dyn_cast<LoadOp>(src.getDefiningOp())) {
+    auto &bld = LF.getRewriter();
+    auto addr = load.getAddr();
+
+    auto oldAlloca = mlir::dyn_cast<AllocaOp>(addr.getDefiningOp());
+    auto alloca = bld.create<AllocaOp>(
+        src.getLoc(), bld.getType<PointerType>(ty), ty,
+        /*name=*/llvm::StringRef(""), oldAlloca.getAlignmentAttr());
+
+    auto tySize = LF.LM.getDataLayout().getTypeStoreSize(ty);
+    createMemCpy(LF, alloca, addr, tySize.getFixedValue());
+
+    auto newLoad = bld.create<LoadOp>(src.getLoc(), alloca.getResult());
+    bld.replaceAllOpUsesWith(load, newLoad);
+
+    return newLoad;
+  }
+
+  cir_cconv_unreachable("NYI");
+}
+
 /// After the calling convention is lowered, an ABI-agnostic type might have to
 /// be loaded back to its ABI-aware couterpart so it may be returned. If they
 /// differ, we have to do a coerced load. A coerced load, which means to load a
@@ -370,7 +395,8 @@ mlir::Value castReturnValue(mlir::Value Src, mlir::Type Ty, LowerFunction &LF) {
 
   auto intTy = mlir::dyn_cast<IntType>(Ty);
   if (intTy && !intTy.isPrimitive())
-    cir_cconv_unreachable("non-primitive types NYI");
+    return createCoercedNonPrimitive(Src, Ty, LF);
+
   llvm::TypeSize DstSize = LF.LM.getDataLayout().getTypeAllocSize(Ty);
 
   // FIXME(cir): Do we need the EnterStructPointerForCoercedAccess routine here?

clang/lib/CIR/Dialect/Transforms/TargetLowering/RecordLayoutBuilder.cpp

@@ -223,6 +223,9 @@ class ItaniumRecordLayoutBuilder {
   /// Initialize record layout for the given record decl.
   void initializeLayout(const Type Ty);
 
+  /// Finalize record layout. Adjust record size based on the alignment.
+  void finishLayout(const StructType D);
+
   uint64_t getDataSizeInBits() const { return DataSize; }
 
   void setDataSize(clang::CharUnits NewSize) {
@@ -243,8 +246,7 @@ void ItaniumRecordLayoutBuilder::layout(const StructType RT) {
   // FIXME(cir): Handle virtual-related layouts.
   cir_cconv_assert(!cir::MissingFeatures::getCXXRecordBases());
 
-  cir_cconv_assert(
-      !cir::MissingFeatures::itaniumRecordLayoutBuilderFinishLayout());
+  finishLayout(RT);
 }
 
 void ItaniumRecordLayoutBuilder::initializeLayout(const mlir::Type Ty) {
@@ -478,6 +480,31 @@ void ItaniumRecordLayoutBuilder::layoutFields(const StructType D) {
   }
 }
 
+void ItaniumRecordLayoutBuilder::finishLayout(const StructType D) {
+  // If we have any remaining field tail padding, include that in the overall
+  // size.
+  setSize(std::max(getSizeInBits(), (uint64_t)Context.toBits(PaddedFieldSize)));
+
+  // Finally, round the size of the record up to the alignment of the
+  // record itself.
+  uint64_t unpaddedSize = getSizeInBits() - UnfilledBitsInLastUnit;
+  uint64_t unpackedSizeInBits =
+      llvm::alignTo(getSizeInBits(), Context.toBits(UnpackedAlignment));
+
+  uint64_t roundedSize = llvm::alignTo(
+      getSizeInBits(),
+      Context.toBits(!Context.getTargetInfo().defaultsToAIXPowerAlignment()
+                         ? Alignment
+                         : PreferredAlignment));
+
+  if (UseExternalLayout) {
+    cir_cconv_unreachable("NYI");
+  }
+
+  // Set the size to the final size.
+  setSize(roundedSize);
+}
+
 void ItaniumRecordLayoutBuilder::UpdateAlignment(
     clang::CharUnits NewAlignment, clang::CharUnits UnpackedNewAlignment,
     clang::CharUnits PreferredNewAlignment) {
@@ -521,13 +548,13 @@ void ItaniumRecordLayoutBuilder::checkFieldPadding(
 
   // Warn if padding was introduced to the struct/class.
   if (!IsUnion && Offset > UnpaddedOffset) {
-    unsigned PadSize = Offset - UnpaddedOffset;
-    // bool InBits = true;
-    if (PadSize % CharBitNum == 0) {
-      PadSize = PadSize / CharBitNum;
-      // InBits = false;
+    unsigned padSize = Offset - UnpaddedOffset;
+    bool inBits = true;
+    if (padSize % CharBitNum == 0) {
+      padSize = padSize / CharBitNum;
+      inBits = false;
     }
-    cir_cconv_assert(cir::MissingFeatures::bitFieldPaddingDiagnostics());
+    cir_cconv_assert(!cir::MissingFeatures::bitFieldPaddingDiagnostics());
   }
   if (isPacked && Offset != UnpackedOffset) {
     HasPackedField = true;

clang/test/CIR/CallConvLowering/AArch64/aarch64-cc-structs.c

@@ -205,3 +205,32 @@ GT_128 call_and_get_gt_128() {
 // LLVM:   store [2 x i64] %[[#ARG]], ptr %[[#V2]], align 8
 // LLVM:   call void @llvm.memcpy.p0.p0.i64(ptr %[[#V1]], ptr %[[#V2]], i64 12, i1 false)
 void passS(S s) {}
+
+typedef struct {
+  uint8_t a;
+  uint16_t b;
+  uint8_t c;
+} S_PAD;
+
+// CHECK: cir.func {{.*@ret_s_pad}}()  -> !u48i
+// CHECK: %[[#V0:]] = cir.alloca !ty_S_PAD, !cir.ptr<!ty_S_PAD>, ["__retval"] {alignment = 2 : i64}
+// CHECK: %[[#V1:]] = cir.load %[[#V0]] : !cir.ptr<!ty_S_PAD>, !ty_S_PAD
+// CHECK: %[[#V2:]] = cir.alloca !u48i, !cir.ptr<!u48i>, [""] {alignment = 2 : i64}
+// CHECK: %[[#V3:]] = cir.cast(bitcast, %[[#V0]] : !cir.ptr<!ty_S_PAD>)
+// CHECK: %[[#V4:]] = cir.cast(bitcast, %[[#V2:]] : !cir.ptr<!u48i>), !cir.ptr<!void>
+// CHECK: %[[#V5:]] = cir.const #cir.int<6> : !u64i
+// CHECK: cir.libc.memcpy %[[#V5]] bytes from %[[#V3]] to %[[#V4]] : !u64i, !cir.ptr<!void>
+// CHECK: %[[#V6:]] = cir.load %[[#V2]] : !cir.ptr<!u48i>
+// CHECK: cir.return %[[#V6]]
+
+// LLVM: i48 @ret_s_pad()
+// LLVM: %[[#V1:]] = alloca %struct.S_PAD, i64 1, align 2
+// LLVM: %[[#V2:]] = load %struct.S_PAD, ptr %[[#V1]], align 2
+// LLVM: %[[#V3:]] = alloca i48, i64 1, align 2
+// LLVM: call void @llvm.memcpy.p0.p0.i64(ptr %[[#V3]], ptr %[[#V1]], i64 6, i1 false)
+// LLVM: %[[#V4:]] = load i48, ptr %[[#V3]]
+// LLVM: ret i48 %[[#V4]]
+S_PAD ret_s_pad() {
+  S_PAD s;
+  return s;
+}